Duncan Eddy, Stanford University; Isaac Ward, Stanford University; Grace Kim, Stanford University; Mykel Kochenderfer, Stanford University
https://doi.org/10.64861/NPSG8113
Keywords: Space Domain Awareness, Uncertainty Modeling, World Models, Machine Learning, Artificial Intelligence, Rendezvous and Proximity Operations, RPO, GPU
Abstract:
This paper introduces a world model-based approach to cooperative and non-cooperative spacecraft proximity operations, demonstrating a learning-based method for spacecraft inspection and rendezvous that combines multi-step planning with calibrated state prediction uncertainty._x000D_
_x000D_
World models are an emerging paradigm in representation learning where agents co-learn joint state-action system dynamics and observation models from offline trajectory data, producing a learned system model. They have demonstrated strong results in robotics and game domains by enabling multi-step reasoning, model-based planning, and quantification of predictive uncertainty—capabilities that are particularly relevant to spacecraft operations where decisions must account for dynamics evolving over orbital timescales._x000D_
_x000D_
Despite their promise, to the best of our knowledge, world models have not been previously applied to spacecraft proximity operations. Key barriers include the lack of GPU-accelerated astrodynamics simulation environments integrated with modern learning pipelines, and limited ability to support the long-horizon planning needed for space applications where fuel-optimal maneuver opportunities occur only at specific orbital locations minutes to hours apart. World models typically require millions of state-action transition pairs for successful training, making efficient simulation over appropriate dynamics regimes essential to generating sufficient training data._x000D_
_x000D_
This paper makes three contributions. First, we introduce an open-source, JAX-based ISS-docking simulation environment supporting massively parallel GPU-based simulation of spacecraft orbit and attitude propagation. Second, we introduce Out of This World Model, a transformer-based world model architecture for space systems. The model combines kinematic and visual observations from relative state measurements and body-fixed cameras using learning-based encoders, producing a latent vector that describes what the satellite is observing at a given time. This latent is fed into a transformer alongside the control action (commanded per-axis thrusts and control torques) executed by the satellite at that time, and over the course of many epochs of training, the model learns to predict the future observation, given the action. We leverage a 1-step flow matching approach to perform this complex, multimodal prediction. In this way, Out of This World Model produces a distribution over future observations in its latent space, allowing for modeling stochastic dynamics and uncertainty. We provide comparisons against DreamerV3-style models that leverage a posterior correction-based approach, and show that our method provides greater predictive performance with fewer trainable parameters and hyperparameters. As an additional benefit, the distributional output provides per-timestep uncertainty estimates over predicted states, which can serve as a basis for runtime anomaly detection and safety monitoring during proximity operations._x000D_
_x000D_
Finally, we apply this approach to the problem of a capsule docking with the International Space Station (ISS). The scenario can be cooperative or non-cooperative, with the agent decision only relying on relative cartesian state, inertial attitude, and a single body-fixed camera. The capsule must autonomously control its orbit and attitude to approach and dock with the ISS at a specified berthing location while respecting keep-out zones. We compare our method against reinforcement learning baselines and demonstrate improvements in sample efficiency and task performance. Furthermore, we show that the world-model based approach has improved generalization properties compared to reinforcement learning methods when tasked with docking at a port not used in the training data. Finally, we show that the learned world model is able to successfully detect anomalies with respect to the training data, correctly identifying when capsules not seen in the training data are visible during the docking approach._x000D_
_x000D_
These results represent a step toward enabling autonomous spacecraft to reason about the consequences of their actions before executing them—a capability with direct applications to on-orbit servicing, active debris removal, and autonomous inspection missions. As the orbital environment grows increasingly complicated and operational tempo demands faster decision-making, methods that combine learned system models with principled uncertainty quantification offer a path toward safer and more capable autonomous proximity operations. We open-source the simulation environments and model architecture to enable further study and exploration of this paradigm._x000D_
Date of Conference: September 15-18, 2026
Track: Conjunction/RPO